COMPUTER VISION-BASED APPROACH FOR MARKERLESS UAV LANDING ZONE IDENTIFICATION

DOI: 10.31673/2412-4338.2025.038701

Abstract

This article presents a computer vision-based approach for autonomous detection of safe landing zones for unmanned aerial vehicles (UAVs) using only images from an onboard camera. YOLOv5s was chosen as the base model, providing a good balance in speed and accuracy of detection with low computational complexity, allowing deployment in the resource-limited environments. The model was trained on a database of an urban environment containing four classes: “landing impossible”, “landing possible”, “person”, and “tree”. To increase the robustness, two data augmentation strategies were proposed that extend the input image processing pipeline at the model level. The first one utilizes the CLAHE, ToGray, or Equalize image augmentations, and another uses RandomBrightnessContrast, RandomShadow, or GaussNoise. To check the adaptability of the proposed models to real-world variations, the evaluation session was conducted
with different confidence thresholds: 25%, 50% and 75%. The results show that the modified models demonstrate a moderate improvement in key metrics. To further optimize results, an additional fine-tuning round was conducted using optimized hyperparameters and the weights from the initial stage. The results of evaluation highlight the efficiency of the
proposed approaches. Finally, based on the research results, the best model was selected for further use. Directions for future research are outlined, focusing on creating autonomous last-mile delivery systems using UAVs to increase the reliability and efficiency of delivery.

Keywords: object detection, image processing, unmanned aerial vehicles, UAV, autonomous landing, landing zone detection, autonomous delivery